Reading your data-retention policy and your rate limits
Who this guide is for: every member of the business — owners, administrators and employees can read both of these disclosures; they need no special permission beyond active membership.
Two read-only disclosure pages tell your business what the platform actually does with its data and its requests:
- Data classes and retention — how long conversations and consent records are kept, and whether your business has its own override of the platform default.
- Rate limits in force — the request limits this service actually enforces, published as resolved numbers.
Both answer with "private, no-store" caching — a declaration about your business is never cached and never shared.
About the numbers. The override windows (conversations 24 months, consent 7 years) and every rate number quoted below come from isolated synthetic test runs behind this guide. No live customer configuration was read, and no live request was limited.
Part 1 — Retention: the default, your override, and no leakage
The retention answer carries the policy in force for your business and says plainly whether it is the platform default or your own override:
- A business with no override reads:
{ "conversationMonths": 12, "consentYears": 3, "isOverride": false, "source": "platform-default" }. - A business with a stored override (24 months / 7 years in the test) reads
{ "conversationMonths": 24, "consentYears": 7, "isOverride": true, "source": "org-override" }— an override always reads as the override, never dressed up as the default. - Reading one business's policy never contaminates another's. In the test, one member read the default policy, a member of the override business read the override, and the first member re-read and got the default again — byte-identical to the first answer.
Behind that read, the storage discipline was proven on the real database: the retention-policies table has forced row-level security, the application uses an ordinary database role, a query pinned to business A sees zero rows of business B's policy, and any attempt to write another business's policy row is refused by the database itself.
Part 2 — Rate limits: the resolved numbers, verbatim
The rate-limits answer publishes the limits the service actually resolved from its configuration — per group, both per-organization and per-IP:
- Three groups, exactly:
auth-heavy,writesandreads— each with per-organization and per-IP limits per minute. - In the test, two separately-configured instances disclosed their own numbers verbatim. The answer is the instance's resolved truth — there is no hand-written constant that could drift from it.
- The answer also names the exempt paths — the health-check routes are the only routes the limiter never counts.
- Configuration itself stays honest: a non-numeric rate value is rejected at load time, not silently coerced.
If you are refused
Both disclosure pages refuse identically:
| What you see | Why | What to do |
|---|---|---|
| 401 AUTH_MISSING | No sign-in was sent | Sign in again |
| 401 AUTH_FORGED | The credential was signed with the wrong secret | Re-authenticate with a valid credential |
| 403 MEMBERSHIP_REQUIRED | You are not a member of the selected business | Switch to a business you belong to |
| 503 DB_UNAVAILABLE | The database is temporarily unavailable | Retry shortly; you will get a clear answer, never a fabricated declaration |
After every refusal the same member's plain read returned success again — refusals are honest stops, never a stuck session.
The preserved failure — why this guide can promise the 503
The test run's first attempt had two real failures (with the database
unavailable, both pages returned a plain 500 instead of the required 503).
That failure is kept, not buried: the failing log is preserved and the fix
(c2b87090) changed only the unavailable-database conversion — no assertion
was weakened. The promise was then proven fixed on a fresh database twice:
when the database is unavailable, both disclosures answer 503
DB_UNAVAILABLE — never a 500, never a fabricated declaration.
Evidence record (kept for honesty). The FULL3747-PG lane executed the
mounted P19 suite against the real buildApp composition — not a
standalone router — on a real ordinary-role (NOSUPERUSER, NOBYPASSRLS)
throwaway PostgreSQL 16 cluster, over real loopback HTTP, on the combined
candidate:
- Suite 13 —
p19-disclosures.build-app-pg.test.ts: first run 5 passed / 2 failed (2026-10-24T23:25:03Z → 23:25:22Z, exit 1,logs/13-p19.tap, failing original preserved verbatim). The two failures were real (plain500 INTERNALinstead of the required 503 envelope with the DB unavailable). - Fix
c2b87090converts the membership lookup's unavailable-DB error to the required503 DB_UNAVAILABLE; assertions unchanged. - Rerun after rebuild: 7/7 passed at 2026-10-24T23:33Z
(
logs/13-p19-rerun1.tap), and a second independent fresh-PG cycle was green again (c2-13-p19.tap; lane ledger "CYCLES 1+2 COMPLETE AND GREEN", 2026-10-24T23:44:30Z). - Receipts in the PG lane worktree
C:/CT-ZC-FULL3747/tools/zcode-integration3723/runtime/FULL3747-PG/logs/(git-ignored); lane ledgerC:/CT-ZCODE-FULL3747/lane-pg-ledger.json.
Standing: combined-candidate, local evidence — the disclosure routes are
mounted in the candidate's app.ts and were exercised through the real
buildApp. Chief acceptance and publication OPEN. A sibling guides
branch (codex/zcode-3456-guides, not part of this guide's base) documents
the UI card journeys for these same disclosures with their own receipts;
this guide documents the mounted API contract the cards read from. Nothing
here claims a live or Main-built deployment.